Home / AI GameChanger / Generative AI & LLMs
✨ Generative AI & LLMs

Fine-Tuning vs RAG vs Prompting — Which to Use

Intermediate ⏱ 6 min read 📘 Lesson 25 of 33

"How do I make the AI know about X / behave like Y?" There are three levers, in increasing order of cost. Pick the cheapest that works — a classic senior-engineer instinct.

1. Prompting (start here — free, instant)

Just describe what you want, give examples in the prompt. Solves a surprising amount. Change behaviour by editing text, not retraining.

2. RAG (for knowledge — cheap, flexible)

Use when the model needs to know facts it wasn't trained on: your docs, recent data, private info. Update by changing documents. See RAG.

3. Fine-tuning (for behaviour/style — expensive, last resort)

Further-train the model on hundreds/thousands of examples. Use to bake in a consistent style, format or skill — not to add knowledge (RAG is better for facts).

Need the model to KNOW your data?          → RAG
Need a consistent STYLE / FORMAT / tone?   → Fine-tune
Need a one-off behaviour change?           → Prompt
Not sure?                                  → Prompt, then RAG, then fine-tune

The decision table

  • "Answer questions about our 500 support articles" → RAG
  • "Always reply in our brand voice as valid JSON" → fine-tune (or a strong system prompt first)
  • "Summarise this text" → prompt
  • "Classify tickets into our 12 internal categories, consistently, at scale" → prompt with examples → fine-tune if accuracy/cost demands

Rule of thumb: 90% of "we need to fine-tune" turns out to be solved by a better prompt + RAG, at a fraction of the cost and effort. Reach for fine-tuning only when you have proven the cheaper levers fall short.